HAM10000 Skin Lesion Classification

This repository contains a two-phase deep learning pipeline for classifying skin lesion images from the HAM10000 dataset. The project uses Supervised Contrastive Learning for feature representation and a fine-tuned classifier with Focal Loss for handling class imbalance.

---

Table of Contents

• Project Overview
• Dataset
• Pipeline
• Requirements
• Usage
• Results
• License

---

Project Overview

Skin cancer detection is a critical problem in medical imaging. This project implements a deep learning approach to classify skin lesions into seven categories:

• akiec
• bcc
• bkl
• df
• mel
• nv
• vasc

The pipeline is divided into two phases:

1. Phase 1: Supervised Contrastive Pretraining

   • Input images → Data augmentation → Encoder (EfficientNetB0) → Projection head (128-D features)
   • Loss: Supervised Contrastive Loss
   • Purpose: Learn robust and discriminative feature representations

2. Phase 2: Fine-tuning Classifier

   • Frozen/partially frozen encoder → MLP classifier → Prediction
   • Loss: Focal Loss (γ=3) to handle class imbalance
   • Optimizer: Adam with low learning rate for fine-tuning

---

Dataset

• HAM10000 Dataset: A large collection of multi-source dermatoscopic images of common pigmented skin lesions.
• Contains images and metadata CSV file.

Folder structure:

HAM10000/
├── HAM10000_images_part_1/
├── HAM10000_images_part_2/
└── HAM10000_metadata.csv

---

Pipeline Overview

Phase 1:
image -> augment -> EfficientNet encoder -> projection (128-D) -> SupCon Loss -> 1792-D embeddings

Phase 2:
encoder (frozen/fine-tuned) -> MLP classifier -> Focal Loss -> predictions

Requirements

torch
torchvision
timm
albumentations
opencv-python
scikit-learn
pandas
numpy

Install all dependencies via:

pip install -r requirements.txt

Usage

1. Clone repository:

git clone https://github.com/username/ham10000-classification.git
cd ham10000-classification

2. Prepare dataset in HAM10000/ folder.

3. Run Phase 1 (Encoder Pretraining):

# Execute your phase 1 code in train_encoder.py or notebook
python train_encoder.py

4. Run Phase 2 (Classifier Fine-tuning):

# Execute classifier training
python train_classifier.py

5. Evaluate model:

# Use your test/validation loader for predictions
python evaluate.py

Results

Validation Metrics:

Class	Precision	Recall	F1-score
bkl	0.71	0.82	0.76
nv	0.97	0.89	0.93
df	0.54	0.65	0.59
mel	0.68	0.74	0.71
vasc	0.95	0.75	0.84
bcc	0.68	0.93	0.79
akiec	0.62	0.72	0.67

Overall Accuracy: 86%
Macro F1-score: 0.75

Author

Md Nahidul Islam
GitHub: leon-dream1